3

I am studying for a test, one section of which will cover the delta method. This problem came from that section:

Let $X\sim N(\mu,n^{-1})$. Find an approximate distribution of $X^2$. (It also asks for exact distribution, but I can do that).

Can you help me get started? I can't see how I can apply delta method or central limit theorem etc. to find some approximation.

I know that $\sqrt{n}(X-\mu) \sim N(0,1)$, but I don't see how that can help either.

bdeonovic
  • 8,507
  • 1
  • 24
  • 49

1 Answers1

2

Let's double check what the delta method is:

Roughly, if there is a sequence of random variables Xn satisfying $$ {\sqrt{n}[X_n-\theta]\,\xrightarrow{D}\,\mathcal{N}(0,\sigma^2)}, $$ where $\theta$ and $\sigma^2$ are finite valued constants and $\xrightarrow{D}$ denotes convergence in distribution, then $$ {\sqrt{n}[g(X_n)-g(\theta)]\,\xrightarrow{D}\,\mathcal{N}(0,\sigma^2[g'(\theta)]^2)} $$ for any function $g$ satisfying the property that $g′(\theta)$ exists and is non-zero valued.

Okay, so showing you have the first thing already is easy right?

So let $g(x) = x^2$ and you're set, aren't you? Just apply the theorem.

Then at the end do the relevant linear transformation (affecting variance and mean) so that you're just talking about what $g(X)$ will approximately be distributed as.

So in short:

i) State the theorem.

ii) Explain/show how the first part is okay

iii) state $g$

iv) apply the theorem

v) infer approximate distribution for $g(X)$

Which is pretty much what you do any time you want to use it.

Glen_b
  • 257,508
  • 32
  • 553
  • 939
  • Well I $\sqrt{n}(X-\mu) \sim N(0,1)$ which is a bit different, but I suppose it is stronger than simply approaching in distribution... – bdeonovic Feb 21 '13 at 03:08
  • Since it's already there, it gets there really fast. – Glen_b Feb 21 '13 at 03:10
  • Yes, I see. I guess it was a silly question then, also quite easy :) – bdeonovic Feb 21 '13 at 03:12
  • Actually this is kind of an interesting problem. I get something like $X^2 \overset{\circ}{\sim} N(\mu^2,4\mu^2/n)$ which is fine as long as $\mu\ne 0$... perhaps using a second order taylor expansion in delta method will be better. – bdeonovic Feb 21 '13 at 03:17
  • You need $\mu$ not close to zero (e.g. many sds - a good deal more than 3 in general); then the approximation is okay. e.g. take a $N(10,1^2)$ and simulate; its square should be approx $N(10^2,(2\times 1\times 10)^2)$... and by my simulations it is. Yes, if you want it to work well in a wider range of circumstances, you may need more terms. – Glen_b Feb 21 '13 at 03:26
  • I think we can come up with a better approximation for when $\mu=0$ :) – bdeonovic Feb 21 '13 at 03:27
  • 1
    aha! Turns out a better approximation when $\mu=0$ is a $\chi_1^2$ random variable, which makes perfect sense since a normal random variable squared is something like a $\chi^2$ random variable. – bdeonovic Feb 21 '13 at 03:37
  • I thought that's what you meant in your earlier comment... – Glen_b Feb 21 '13 at 03:41